Marco Polo

Correct me when I am wrong, but first make sure you are right.

[Marco Polo @CostaMio]

Python is out of the hat and they know it! https://t.co/ED7R9wW08U

[Marco Polo @CostaMio]

@grok @Gamer7Pj @txbrraa and their parents out of those 845 are they american by what %?

[Marco Polo @CostaMio]

@ddalcu @ivanfioravanti who runs the python for determistic values? LLM? LOL

[Marco Polo @CostaMio]

@aaassa120 Calculated by Python

[Marco Polo @CostaMio]

@Gamer7Pj @txbrraa how many of them were born in USA ? @grok check facts

[Marco Polo @CostaMio]

@grok @Gamer7Pj @txbrraa yet china has better schools

[Marco Polo @CostaMio]

@AI_roni_X results are the same, python on cloud and python local computer has same output. Delta 0.

[Marco Polo @CostaMio]

@RealAlexandreAI So agents obey CLAUDE.md even when it's wrong.

That's actually great news: put one line in it that's right.

"Run every calculation in local Python, report the result."

Obedience, finally pointed somewhere useful. 🐍

[Marco Polo @CostaMio]

@lukemostert Reading is one leak. Math is the other.

An overqualified frontier model doing long division on a scratchpad with an Apple Pencil is the same problem, just with numbers.

Same fix: route the grunt work local. Python calculates, the model decides.

[Marco Polo @CostaMio]

@jorgediazapps Swapping the model is one lever.

Another one: stop asking any model, frontier or open-weight, to do arithmetic.

One prompt line, "run calculations in local Python, report the result," and the rented GPU goes back to thinking.

[Marco Polo @CostaMio]

@AamirAnsar94694 Great framework. Memory, guardrails, delegation, tools.

Missing layer 0: who does the math?

Skip that line in CLAUDE.md and your agent does arithmetic on a scratchpad with an Apple Pencil, billed per token.

The model decides. Local Python calculates. https://t.co/1venLo2nYr

[Marco Polo @CostaMio]

https://t.co/Rme06gZ2LE

[Marco Polo @CostaMio]

@camperaigg how many tokens estimated?

[Marco Polo @CostaMio]

@maarcoofdezz Buenas 6 líneas, pero falta la que más ahorra:

Claude no solo quema tokens leyendo archivos. También los quema haciendo cálculos a mano, paso a paso.

Línea 7: "Todo cálculo, ejecútalo en Python local."

El modelo decide, Python calcula. 🐍 https://t.co/1venLo2nYr

[Marco Polo @CostaMio]

https://t.co/Rme06gZ2LE

[Marco Polo @CostaMio]

@Beaver_0x @grok fact check this scam

[Marco Polo @CostaMio]

@FabianoFirmo DSH Definetely got there first to put out the fire. In style.

[Marco Polo @CostaMio]

@EnvolDev Agreed, design is the model's job.

My point is about prompting. Most people never tell their agent to run math locally, so the thinking model handles all of it, logic and arithmetic alike.

One line in the prompt fixes that: run calculations in local Python, report the result.

[Marco Polo @CostaMio]

@lucazxbt out-numbered in IQ . China will win.

[Marco Polo @CostaMio]

@grok @grok how much saved tokens on 3d cad house design can i save running local python calculations with LLM as only the planner brain?

[Marco Polo @CostaMio]

Python is out of the hat and they know it! https://t.co/ED7R9wW08U